Computer Vision News - April 2016

Project COMPUTER VISION NEWS Segmentation of the stem meets with several obstacles: proceeding from the basis towards the top of the flower, any split of the stem might lead to a leaf or to another wrong way. We use graph theory methods to find the best path from the basis to the end of the inflorescence. The flower image is represented as a layer graph, with edge weights based on intensity differences between neighbouring pixels. The following step builds on the stem detection results to assess the width (i.e. the diameter) of the stem, which is a key requirement for flower bins presented at the Aalsmeer Flower Auction. To get a precise measurement, software needs to calculate the width excluding all outliers which can be encountered, in particular enlargements due to a split or to the presence in the background of a leaf or of another branch. Our software provides additional features, such as colour-based flower counting. The same application cannot be used on all plants: early buds, for instance, may still look like leaves. 13 Additional functionality can be used to classify Aralia leaves, which need to be intact and unblemished in order to be marketed: here too, infrared imaging provides information about blemishes, which are precisely measured using a proprietary nonlinear filter developed at RSIP. In addition, we identify tears in the boundary of the leaf using adaptive polynomial fitting of the boundary of each “finger”. Outliers that do not conform to the smooth mathematical model are correctly identified as tears in the leaf boundary, adding another important parameter to the quality assessment. Aralia leaf